Speech Recognition
CaptionCall, Sorenson Communication
As a scientist on the Machine Learning and Speech Processing team at CaptionCall, I developed algorithms and implemented machine learning and natural language processing techniques to improve automatic speech recognition for telephone captioning, resulting in deployment and three patents. My contributions included:
Successfully developed novel techniques for training n-gram language models on the fly without saving transcriptions, including techniques that make use of interpolation and neural net text sampling.                                                                                                            Â
Implemented natural language processing and deep learning methods to estimate transcription quality with an average mean absolute error of 1.6% (more than ten times smaller than state-of-the-art estimators at the time).Â
Collaborated in implementing real-time fusion of multiple transcripts for improved accuracy. Contributed quality estimation of transcripts using support vector machines for enhanced voting and tie-breaking decisions, improving accuracy by 5% on average.
Training Speech Recognition Systems Using Word SequencesÂ
US Patent 10,388,272 and 10,672,383 (2020)
Transcription Generation from Multiple Speech Recognition SystemsÂ
US Patent 10,573,312 and 10,971,153 (2020)